Predicting the Past: Google DeepMind branches Gemini on its epigraphy models

Google DeepMind launches Predicting the Past, a Gemini tool using Ithaca and Aeneas models to restore, date, and locate ancient Greek and Latin inscriptions.

Restoring a mutilated ancient inscription, dating it, locating it: these are all puzzles that epigraphists solve on a case-by-case basis. Google DeepMind proposes to conduct this work in natural language with Predicting the Past, a skill for its Google Antigravity environment. The principle consists of anchoring Gemini in the outputs of two already proven in-house models, Ithaca (2022) and Aeneas (2025), specialized in the restoration, dating, and attribution of Greek and Latin inscriptions. The historian attributes, reconstructs, and analyzes a text as if in an exchange with a colleague, without writing code.

The approach addresses three obstacles identified with the community: producing visualizations specific to each inscription, cross-referencing multiple texts to identify large-scale patterns, and keeping the language model anchored to verifiable evidence. Conducted with epigraphist Thea Sommerschield (Durham University), three cases illustrate the range: the attribution of a Latin curse tablet from Bath, where Aeneas places the inscription within the ranges accepted by historians while explaining its reasoning; the mapping of a Rhenish cult from an altar in Mainz, to trace the circulation of religious practices in the Empire; and the mass processing of the oracular tablets of Dodona, whose model reconstructs the community of visitors. The Ithaca and Aeneas models are open source, and an online access platform already exists.